TL;DR
This survey explores how neural networks (NNs) are revolutionizing molecular communication (MC) for the Internet of Bio-Nano Things (IoBNT). By leveraging NNs, researchers are developing more robust, adaptive, and explainable communication strategies for nanoscale devices, paving the way for advanced healthcare applications.
What Problem Does This Paper Address?
The vision of the Internet of Bio-Nano Things (IoBNT) involves a network of tiny devices communicating within the human body to monitor and treat diseases. However, the unique characteristics of molecular communication (MC) channels – their stochastic, dynamic, and often analytically intractable nature – pose significant challenges for traditional communication design. Purely model-based approaches struggle to cope with these complexities. This paper addresses the need for data-driven communication strategies, specifically focusing on how neural networks (NNs) can enable reliable, explainable, and deployable communication in these nanoscale environments.
Methods
The paper surveys the application of various NN architectures to different layers of the MC communication stack, including:
- Channel Estimation: NNs are used to learn the complex impulse response (CIR) of MC channels, which is often infeasible to model analytically.
- Synchronization: Reinforcement learning (RL) agents and deep neural networks (DNNs) are employed to adapt to dynamic channel conditions and synchronize communication timing.
- Detection: Feedforward NNs, RNNs, CNNs, and transformers are used to decode incoming symbols, overcoming the challenges of unknown channel distributions and parameters.
- Autoencoders (AEC): These are used for end-to-end learning, optimizing both the transmitter and receiver components of the MC system simultaneously, mitigating channel modeling challenges.
- Higher Layers: NNs are also being explored for resource allocation (MAC layer) and localization applications.
The survey also delves into the feasibility of implementing these NNs at the nanoscale using biocomputing techniques (digital, microfluidic, and chemical) and discusses explainable AI (XAI) methods to understand NN decision-making.
Datasets
The paper discusses the importance of datasets for training NNs in MC and categorizes them into:
- Synthetic Datasets: Generated through simulations (e.g., using OpenFOAM, ANSYS Fluent, COMSOL Multiphysics, nanoNS3) that mimic various MC channel environments like pipe-like and free-diffusion channels.
- Experimental Datasets: Collected from real-world testbeds, including those using fluorescent particles, bacteria, or salinity changes as information carriers.
The authors highlight the need for publicly available, well-documented, and representative datasets to facilitate reproducible research. They also review various dataset generation tools and provide a traffic-light system to evaluate dataset accessibility and usability.
Key Insights and Ideas
- NNs as a Paradigm Shift: Unlike traditional wireless communication where ML was introduced later, NNs are foundational to IoBNT/MC design from its inception, enabling adaptive solutions for inherently complex channels.
- Biocompatibility is Key: For IoBNT, NNs must be "biocompatible," moving beyond standard digital implementations to low-energy modules that can be hosted within biological systems.
- Explainability is Non-Negotiable: In healthcare applications, "black-box" NNs are insufficient. XAI techniques are crucial for building trust and understanding how NNs make decisions in biological contexts.
- End-to-End Learning for Optimization: Autoencoders offer a powerful approach to optimize the entire MC system, bypassing the need for explicit channel modeling and potentially achieving higher performance.
- The Data Bottleneck: The availability and quality of datasets are critical for training robust NN models. Standardization and accessibility of data are major challenges.
Key Results
- NN-based detectors have achieved Bit Error Rates (BERs) of less than $10^{-4}$ in testbed environments.
- BiRNN architectures demonstrate similar performance to other complex architectures with fewer learnable parameters, offering a good trade-off between performance and complexity.
- XAI methods like LIME and ICE have shown that trained NNs can mirror the behavior of traditional detectors (e.g., peak and slope detectors) in certain MC scenarios, providing a level of interpretability.
- Physics-informed neural networks (PINNs) are emerging as a promising direction for integrating domain knowledge, potentially leading to better generalization and interpretability.
Limitations and Future Work
- Limited Nanoscale Implementations: While research explores biocompatible substrates, practical deployment of complex NNs at the nanoscale remains a significant challenge due to resource constraints.
- Lack of Standardized Configurations: Hyperparameter tuning for NN models in MC is largely empirical, with a lack of standardized approaches across different scenarios.
- Scarcity of Upper-Layer Research: Most research focuses on the PHY layer; more work is needed for MAC and application layers.
- Need for More Realistic Scenarios: Many current NN models are tested in simplified geometries. Future work should target more complex and realistic environments, including those found in the human circulatory system.
- Online and Adaptive Training: Most NNs are trained offline. Developing online and adaptive training strategies is crucial for dynamic MC environments.
- Unified Framework: A standardized framework for designing and implementing biological or chemical NNs is needed.
Summary updated